It sounds like we've found the root cause of most of the speed issue.
Setting --multi-thread-chunk-size 4G should make the code do more or less the same thing as it used to.
I think increasing --multi-thread-streams and decreasing --multi-thread-chunk-size should make it run faster. You could try --multi-thread-chunk-size 1G and --multi-thread-streams 16 and see what difference that makes. You should find just using --multi-thread-streams 16 on its own works quite well too.